In partnership with
Proteins move among shapes as they bind, release partners, and respond to their surroundings. A single structure can describe one observed state, but it may not represent the full motion relevant to a biological process. That distinction matters because artificial intelligence models often return a strong structural prediction from a sequence. Accuracy for one shape does not automatically explain how a protein changes over time.
A fast AI answer isn't always a coordinated one. When AI resolves B2B issues without looping in the teams responsible for them, confidence outpaces alignment. A new Harvard Business Review Analytic Services briefing paper, sponsored by Front, looks at closing that gap.
AlphaFold 3 predicts structures of proteins and other biomolecules from sequence information. Its published research showed accurate predictions of molecular interactions, but later experimental work identified a limit in representing heterogeneous proteins. A June 2026 study in Nature Biotechnology tested a way to guide AlphaFold 3 with laboratory evidence. The aim was to produce a small set of structures whose combined measurements better matched what experiments observed.
The method accepts evidence from nuclear magnetic resonance, X-ray crystallography, and cryogenic electron microscopy. These methods measure different properties. NMR can constrain distances or describe local motion. X-ray data provide electron density within a crystal. Cryogenic electron microscopy supplies density maps for imaged particles. The researchers combined suitable observations with the model’s predictions rather than treating the sequence-based result as a complete answer.
The Agentic Engineering Playbook top engineers are using
Most engineers are stuck babysitting AI one prompt at a time. A handful have figured out how to build agents that plan, execute, and self-correct on their own. Sign up for The Code and get The Complete Playbook for Agentic Engineering. Claude Code, subagents, MCP, all of it.
The study reports that guided structural ensembles matched experimental averages and reduced violations of NMR distance constraints compared with traditionally resolved structures. The method also identified alternate conformations visible in electron density but absent from deposited models. These results show how measurements can add information that a single predicted structure may miss. They do not show that every protein requires the same evidence or that every generated ensemble is correct.
A separate 2026 preprint examined four multi-state proteins with AlphaFold 3, Boltz-2, Chai-1, and BioEmu. The models often favored a dominant state already represented in a structural database. Small-molecule ligands had weak or inconsistent effects in that sample, while larger protein partners produced clearer changes. Because the benchmark used four proteins and remains a preprint, it cannot describe performance across all proteins or laboratory settings.
Together, the studies point toward a different role for structural prediction. A model can supply a useful starting hypothesis, while experimental data test whether that hypothesis fits observed conditions. Researchers can compare the predicted shape with distance constraints, density maps, and evidence of motion. Disagreement then becomes informative: it can signal a missing state, a poor fit, or a limitation in the available measurements.
The distinction matters for drug research, but it does not settle questions about treatment outcomes. A structure that fits an experiment does not prove that a molecule will bind selectively, work in living cells, or become a safe medicine. Those conclusions require separate experiments and later stages of development. Even a better structural ensemble remains a model of evidence, not a substitute for biological testing.
There are practical limits as well. Guidance depends on the quality and coverage of experimental inputs, and different methods observe different parts of a system. The June study describes structural hypotheses constrained by supplied data; it does not claim calibrated probabilities for every molecular state. The independent preprint also shows why broader testing matters. More proteins, conditions, and laboratories are needed before general performance claims are justified.
For readers following scientific AI, the key development is the link between prediction and measurement. The useful question is no longer only whether a model can draw a plausible structure. It is whether the proposed structures explain independent observations and reveal where uncertainty remains. That standard keeps computational results connected to laboratory evidence, while leaving biological interpretation open to further testing. This is a research direction to follow as methods and benchmarks expand.
Your AI prompts are leaving out 80% of what you're thinking.
When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning (constraints, edge cases, examples, tone) and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.
Used by teams at Microsoft, Amazon, and Notion. Try Wispr Flow free on Mac, Windows, iPhone, and Android.
